Transformer Reranker

rankify.models.transformer_ranker

BaseRanking

Bases: ABC

An abstract base class for implementing different ranking models.

This class defines the interface for all ranking models, ensuring that all subclasses implement the required methods.

Attributes:

Name	Type	Description
method	str	The name of the ranking method.
model_name	str	The name of the model being used for ranking.
api_key	str	An optional API key for accessing remote models or services.

Source code in rankify/models/base.py

class BaseRanking(ABC):
    """
    An abstract base class for implementing different ranking models.

    This class defines the interface for all ranking models, ensuring that all subclasses implement the required methods.

    Attributes:
        method (str): The name of the ranking method.
        model_name (str): The name of the model being used for ranking.
        api_key (str, optional): An optional API key for accessing remote models or services.
    """

    @abstractmethod
    def __init__(self, method: str= None, model_name: str= None, api_key: str= None, **kwargs) ->None:
        """
        Initializes the base ranking model.

        Args:
            method (str, optional): The name of the ranking method. Defaults to None.
            model_name (str, optional): The name of the model being used for ranking. Defaults to None.
            api_key (str, optional): An optional API key for accessing remote models or services. Defaults to None.

        Example:
            ```python
            class MyRanking(BaseRanking):
                def __init__(self, method, model_name):
                    super().__init__(method, model_name)
            ```
        """
        pass

    @abstractmethod
    def rank(self, documents: list[Document] ):
        """
        Abstract method to rank a list of documents.

        Args:
            documents (list[Document]): A list of Document instances that need to be ranked.

        Raises:
            NotImplementedError: This method must be implemented by subclasses.

        Example:
            ```python
            class MyRanking(BaseRanking):
                def __init__(self, method, model_name):
                    super().__init__(method, model_name)

                def rank(self, documents):
                    # Ranking implementation here
                    pass
            ```
        """
        pass

__init__(method=None, model_name=None, api_key=None, **kwargs) abstractmethod

Initializes the base ranking model.

Parameters:

Name	Type	Description	Default
method	str	The name of the ranking method. Defaults to None.	None
model_name	str	The name of the model being used for ranking. Defaults to None.	None
api_key	str	An optional API key for accessing remote models or services. Defaults to None.	None

Example

class MyRanking(BaseRanking):
    def __init__(self, method, model_name):
        super().__init__(method, model_name)

Source code in rankify/models/base.py

@abstractmethod
def __init__(self, method: str= None, model_name: str= None, api_key: str= None, **kwargs) ->None:
    """
    Initializes the base ranking model.

    Args:
        method (str, optional): The name of the ranking method. Defaults to None.
        model_name (str, optional): The name of the model being used for ranking. Defaults to None.
        api_key (str, optional): An optional API key for accessing remote models or services. Defaults to None.

    Example:
        ```python
        class MyRanking(BaseRanking):
            def __init__(self, method, model_name):
                super().__init__(method, model_name)
        ```
    """
    pass

rank(documents) abstractmethod

Abstract method to rank a list of documents.

Parameters:

Name	Type	Description	Default
documents	list[Document]	A list of Document instances that need to be ranked.	required

Raises:

Type	Description
NotImplementedError	This method must be implemented by subclasses.

Example

class MyRanking(BaseRanking):
    def __init__(self, method, model_name):
        super().__init__(method, model_name)

    def rank(self, documents):
        # Ranking implementation here
        pass

Source code in rankify/models/base.py

@abstractmethod
def rank(self, documents: list[Document] ):
    """
    Abstract method to rank a list of documents.

    Args:
        documents (list[Document]): A list of Document instances that need to be ranked.

    Raises:
        NotImplementedError: This method must be implemented by subclasses.

    Example:
        ```python
        class MyRanking(BaseRanking):
            def __init__(self, method, model_name):
                super().__init__(method, model_name)

            def rank(self, documents):
                # Ranking implementation here
                pass
        ```
    """
    pass

Document

Represents a document consisting of a question, answers, and contexts.

Attributes:

Name	Type	Description
question	Question	The question associated with the document.
answers	Answer	The answers to the question.
contexts	list[Context]	A list of related contexts.
reorder_contexts	list[Context] or None	A reordered list of contexts based on relevance.

Source code in rankify/dataset/dataset.py

class Document:
    """
    Represents a document consisting of a question, answers, and contexts.

    Attributes:
        question (Question): The question associated with the document.
        answers (Answer): The answers to the question.
        contexts (list[Context]): A list of related contexts.
        reorder_contexts (list[Context] or None): A reordered list of contexts based on relevance.
    """
    def __init__(self, question: Question, answers: Answer, contexts: list = None , id: int = None) -> None:
        """
        Initializes a Document instance.

        Args:
            question (Question): The question associated with the document.
            answers (Answer): The answers to the question.
            contexts (list[Context], optional): A list of contexts related to the question.

        Example:
            ```python
            q = Question("What is the capital of France?")
            a = Answer(["Paris"])
            c1 = Context(score=0.9, has_answer=True, id=1, title="Paris", text="The capital of France is Paris.")
            c2 = Context(score=0.5, has_answer=False, id=2, title="Berlin", text="Berlin is the capital of Germany.")
            d = Document(question=q, answers=a, contexts=[c1, c2])
            print(d)
            ```
        """
        self.question: Question = question
        self.answers: Answer = answers
        self.contexts: List[Context] = contexts
        self.reorder_contexts: List[Context] = None
        self.id = str(id) 

    @classmethod
    def from_dict(cls, data: dict,n_docs:int=100) -> 'Document':
        """
        Creates a Document instance from a dictionary.

        Args:
            data (dict): A dictionary containing the question, answers, and contexts.
            n_docs (int, optional): The number of contexts to include. Defaults to 100.

        Returns:
            Document: A new Document instance.

        Example:
            ```python
            data = {
                "question": "What is the capital of France?",
                "answers": ["Paris"],
                "ctxs": [
                    {"score": 0.9, "has_answer": True, "id": 1, "title": "Paris", "text": "The capital of France is Paris."},
                    {"score": 0.5, "has_answer": False, "id": 2, "title": "Berlin", "text": "Berlin is the capital of Germany."}
                ]
            }
            d = Document.from_dict(data)
            print(d.question)
            ```
        """
        question = Question(data["question"])
        if "answers" in data:
            answers = Answer(data["answers"])
        else:
            answers =Answer('')

        if "query_id" in data:
            id = data["query_id"]
        else:
            id = None
        contexts = [Context(**ctx) for ctx in data["ctxs"][:n_docs]]
        return cls(question, answers, contexts, id=id)

    def to_dict(self) -> Dict[str, Optional[object]]:
        """
        Converts the document into a dictionary representation.

        Returns:
            dict: A dictionary containing the question, answers, and contexts.
        """
        return {
            "question": self.question.question,
            "answers": self.answers.answers,
            "contexts": [ctx.to_dict() for ctx in self.contexts]
        }
    def to_dict_reoreder(self) -> Dict[str,Optional[object]]:
        return {
            "question" : self.question.question,
            "answers" : self.answers.answers,
            "contexts" : [ctx.to_dict() for ctx in self.reorder_contexts]
        }
    def __str__(self) -> str:
        """
        Returns a string representation of the Document instance.

        Returns:
            str: The formatted document information.

        Example:
            ```python
            d = Document(Question("What is the capital of France?"), Answer(["Paris"]))
            print(d)
            ```
        """
        contexts_str = "\n\n".join([str(ctx) for ctx in self.contexts])
        reorder_contexts_str= ''
        if self.reorder_contexts is not None:
            reorder_contexts_str = "\n\n".join([str(ctx) for ctx in self.reorder_contexts])
        return f"{self.question}\n\n{self.answers}\n\nContext: \n\n{contexts_str}\nReorder contexts: \n\n{reorder_contexts_str}"

__init__(question, answers, contexts=None, id=None)

Initializes a Document instance.

Parameters:

Name	Type	Description	Default
question	Question	The question associated with the document.	required
answers	Answer	The answers to the question.	required
contexts	list[Context]	A list of contexts related to the question.	None

Example

q = Question("What is the capital of France?")
a = Answer(["Paris"])
c1 = Context(score=0.9, has_answer=True, id=1, title="Paris", text="The capital of France is Paris.")
c2 = Context(score=0.5, has_answer=False, id=2, title="Berlin", text="Berlin is the capital of Germany.")
d = Document(question=q, answers=a, contexts=[c1, c2])
print(d)

Source code in rankify/dataset/dataset.py

def __init__(self, question: Question, answers: Answer, contexts: list = None , id: int = None) -> None:
    """
    Initializes a Document instance.

    Args:
        question (Question): The question associated with the document.
        answers (Answer): The answers to the question.
        contexts (list[Context], optional): A list of contexts related to the question.

    Example:
        ```python
        q = Question("What is the capital of France?")
        a = Answer(["Paris"])
        c1 = Context(score=0.9, has_answer=True, id=1, title="Paris", text="The capital of France is Paris.")
        c2 = Context(score=0.5, has_answer=False, id=2, title="Berlin", text="Berlin is the capital of Germany.")
        d = Document(question=q, answers=a, contexts=[c1, c2])
        print(d)
        ```
    """
    self.question: Question = question
    self.answers: Answer = answers
    self.contexts: List[Context] = contexts
    self.reorder_contexts: List[Context] = None
    self.id = str(id) 

from_dict(data, n_docs=100) classmethod

Creates a Document instance from a dictionary.

Parameters:

Name	Type	Description	Default
data	dict	A dictionary containing the question, answers, and contexts.	required
n_docs	int	The number of contexts to include. Defaults to 100.	100

Returns:

Name	Type	Description
Document	Document	A new Document instance.

Example

data = {
    "question": "What is the capital of France?",
    "answers": ["Paris"],
    "ctxs": [
        {"score": 0.9, "has_answer": True, "id": 1, "title": "Paris", "text": "The capital of France is Paris."},
        {"score": 0.5, "has_answer": False, "id": 2, "title": "Berlin", "text": "Berlin is the capital of Germany."}
    ]
}
d = Document.from_dict(data)
print(d.question)

Source code in rankify/dataset/dataset.py

@classmethod
def from_dict(cls, data: dict,n_docs:int=100) -> 'Document':
    """
    Creates a Document instance from a dictionary.

    Args:
        data (dict): A dictionary containing the question, answers, and contexts.
        n_docs (int, optional): The number of contexts to include. Defaults to 100.

    Returns:
        Document: A new Document instance.

    Example:
        ```python
        data = {
            "question": "What is the capital of France?",
            "answers": ["Paris"],
            "ctxs": [
                {"score": 0.9, "has_answer": True, "id": 1, "title": "Paris", "text": "The capital of France is Paris."},
                {"score": 0.5, "has_answer": False, "id": 2, "title": "Berlin", "text": "Berlin is the capital of Germany."}
            ]
        }
        d = Document.from_dict(data)
        print(d.question)
        ```
    """
    question = Question(data["question"])
    if "answers" in data:
        answers = Answer(data["answers"])
    else:
        answers =Answer('')

    if "query_id" in data:
        id = data["query_id"]
    else:
        id = None
    contexts = [Context(**ctx) for ctx in data["ctxs"][:n_docs]]
    return cls(question, answers, contexts, id=id)

to_dict()

Converts the document into a dictionary representation.

Returns:

Name	Type	Description
dict	Dict[str, Optional[object]]	A dictionary containing the question, answers, and contexts.

Source code in rankify/dataset/dataset.py

def to_dict(self) -> Dict[str, Optional[object]]:
    """
    Converts the document into a dictionary representation.

    Returns:
        dict: A dictionary containing the question, answers, and contexts.
    """
    return {
        "question": self.question.question,
        "answers": self.answers.answers,
        "contexts": [ctx.to_dict() for ctx in self.contexts]
    }

__str__()

Returns a string representation of the Document instance.

Returns:

Name	Type	Description
str	str	The formatted document information.

Example

d = Document(Question("What is the capital of France?"), Answer(["Paris"]))
print(d)

Source code in rankify/dataset/dataset.py

def __str__(self) -> str:
    """
    Returns a string representation of the Document instance.

    Returns:
        str: The formatted document information.

    Example:
        ```python
        d = Document(Question("What is the capital of France?"), Answer(["Paris"]))
        print(d)
        ```
    """
    contexts_str = "\n\n".join([str(ctx) for ctx in self.contexts])
    reorder_contexts_str= ''
    if self.reorder_contexts is not None:
        reorder_contexts_str = "\n\n".join([str(ctx) for ctx in self.reorder_contexts])
    return f"{self.question}\n\n{self.answers}\n\nContext: \n\n{contexts_str}\nReorder contexts: \n\n{reorder_contexts_str}"

Context

Represents a context with metadata such as score and title.

Attributes:

Name	Type	Description
score	float	The relevance score of the context.
has_answer	bool	Whether the context contains an answer.
id	int	The identifier of the context.
title	str	The title of the context.
text	str	The text of the context.

Source code in rankify/dataset/dataset.py

class Context:
    """
    Represents a context with metadata such as score and title.

    Attributes:
        score (float, optional): The relevance score of the context.
        has_answer (bool, optional): Whether the context contains an answer.
        id (int, optional): The identifier of the context.
        title (str, optional): The title of the context.
        text (str, optional): The text of the context.
    """
    def __init__(self, score: float=None, has_answer: bool=None, id: str=None, title: str=None, text: str=None)-> None:
        """
        Initializes a Context instance.

        Args:
            score (float, optional): The relevance score.
            has_answer (bool, optional): Whether the context contains an answer.
            id (int, optional): The identifier of the context.
            title (str, optional): The title of the context.
            text (str, optional): The text of the context.

        Example:
            ```python
            c = Context(score=0.9, has_answer=True, id=1, title="Paris", text="The capital of France is Paris.")
            print(c)
            ```
        """
        self.score: Optional[float] = score
        self.has_answer: Optional[bool] = has_answer
        self.id: Optional[str] = id
        self.title: Optional[str] = title
        self.text: Optional[str] = text

    def to_dict(self, save_text: bool=False) -> Dict[str, Optional[object]]:

        """
        Converts the Context instance to a dictionary.

        Args:
            save_text (bool): Whether to include text in the output dictionary.

        Returns:
            dict: The context data.

        Example:
            ```python
            c = Context(score=0.9, has_answer=True, id=1, title="Paris", text="The capital of France is Paris.")
            print(c.to_dict())
            ```
        """
        context_dict = {
            "score": float(self.score) if self.score is not None else None,
            "has_answer": self.has_answer,
            "id": self.id,
            }

        # Include 'text' only if save_text is True
        if save_text:
            context_dict["text"] = self.text
            context_dict["title"] =  self.title

        return context_dict
    def __str__(self) -> str:
        """
        Returns a string representation of the Context instance.

        Returns:
            str: The formatted context.

        Example:
            ```python
            c = Context(score=0.9, has_answer=True, id=1, title="Paris", text="The capital of France is Paris.")
            print(str(c))
            ```
        """
        return f"ID: {self.id}\nHas Answer: {self.has_answer}\nTitle: {self.title}\nText: {self.text}\nScore: {self.score}"

__init__(score=None, has_answer=None, id=None, title=None, text=None)

Initializes a Context instance.

Parameters:

Name	Type	Description	Default
score	float	The relevance score.	None
has_answer	bool	Whether the context contains an answer.	None
id	int	The identifier of the context.	None
title	str	The title of the context.	None
text	str	The text of the context.	None

Example

c = Context(score=0.9, has_answer=True, id=1, title="Paris", text="The capital of France is Paris.")
print(c)

Source code in rankify/dataset/dataset.py

def __init__(self, score: float=None, has_answer: bool=None, id: str=None, title: str=None, text: str=None)-> None:
    """
    Initializes a Context instance.

    Args:
        score (float, optional): The relevance score.
        has_answer (bool, optional): Whether the context contains an answer.
        id (int, optional): The identifier of the context.
        title (str, optional): The title of the context.
        text (str, optional): The text of the context.

    Example:
        ```python
        c = Context(score=0.9, has_answer=True, id=1, title="Paris", text="The capital of France is Paris.")
        print(c)
        ```
    """
    self.score: Optional[float] = score
    self.has_answer: Optional[bool] = has_answer
    self.id: Optional[str] = id
    self.title: Optional[str] = title
    self.text: Optional[str] = text

to_dict(save_text=False)

Converts the Context instance to a dictionary.

Parameters:

Name	Type	Description	Default
save_text	bool	Whether to include text in the output dictionary.	False

Returns:

Name	Type	Description
dict	Dict[str, Optional[object]]	The context data.

Example

c = Context(score=0.9, has_answer=True, id=1, title="Paris", text="The capital of France is Paris.")
print(c.to_dict())

Source code in rankify/dataset/dataset.py

def to_dict(self, save_text: bool=False) -> Dict[str, Optional[object]]:

    """
    Converts the Context instance to a dictionary.

    Args:
        save_text (bool): Whether to include text in the output dictionary.

    Returns:
        dict: The context data.

    Example:
        ```python
        c = Context(score=0.9, has_answer=True, id=1, title="Paris", text="The capital of France is Paris.")
        print(c.to_dict())
        ```
    """
    context_dict = {
        "score": float(self.score) if self.score is not None else None,
        "has_answer": self.has_answer,
        "id": self.id,
        }

    # Include 'text' only if save_text is True
    if save_text:
        context_dict["text"] = self.text
        context_dict["title"] =  self.title

    return context_dict

__str__()

Returns a string representation of the Context instance.

Returns:

Name	Type	Description
str	str	The formatted context.

Example

c = Context(score=0.9, has_answer=True, id=1, title="Paris", text="The capital of France is Paris.")
print(str(c))

Source code in rankify/dataset/dataset.py

def __str__(self) -> str:
    """
    Returns a string representation of the Context instance.

    Returns:
        str: The formatted context.

    Example:
        ```python
        c = Context(score=0.9, has_answer=True, id=1, title="Paris", text="The capital of France is Paris.")
        print(str(c))
        ```
    """
    return f"ID: {self.id}\nHas Answer: {self.has_answer}\nTitle: {self.title}\nText: {self.text}\nScore: {self.score}"

TransformerRanker

Bases: BaseRanking

Implements TransformerRanker, a general pretrained transformer-based reranking model.

References

• MixedBread AI (2024): Reranking Overview. Paper
• Xiao et al. (2024): BGE-Reranker: A Packed Resource for General Chinese Embeddings. Paper
• Günther et al. (2023): Jina Embeddings: A Novel Set of High-Performance Sentence Embedding Models. Paper
• mGTE (2024): Generalized Long-Context Text Representation for Multilingual Retrieval. Paper
• Martin et al. (2019): CamemBERT: A Tasty French Language Model. Paper
• BCEmbedding (2023): Bilingual and Cross-lingual Embeddings for RAG. Paper

Attributes:

Name	Type	Description
method	str	The name of the reranking method.
model_name	str	The name or path to the pretrained reranking model.
device	device	The computation device (CPU/GPU).
dtype	dtype	The data type for model inference.
tokenizer	AutoTokenizer	The tokenizer for encoding queries and documents.
batch_size	int	The batch size for efficient inference.

Example

from rankify.dataset.dataset import Document, Question, Context
from rankify.models.reranking import Reranking

# Define a query and contexts
question = Question("What are the benefits of deep learning?")
contexts = [
    Context(text="Deep learning allows models to extract hierarchical representations.", id=0),
    Context(text="Machine learning includes both supervised and unsupervised learning.", id=1),
    Context(text="Neural networks are a fundamental component of deep learning.", id=2),
]
document = Document(question=question, contexts=contexts)

# Initialize Transformer Ranker (e.g., BGE-Reranker)
model = Reranking(method='transformer_ranker', model_name='mxbai-rerank-xsmall')
model.rank([document])

# Print reordered contexts
print("Reordered Contexts:")
for context in document.reorder_contexts:
    print(context.text)

Source code in rankify/models/transformer_ranker.py

class TransformerRanker(BaseRanking):
    """
    Implements **TransformerRanker**, a general **pretrained transformer-based** reranking model.

    References:
        - **MixedBread AI (2024)**: *Reranking Overview*. [Paper](https://www.mixedbread.ai/docs/reranking/overview)
        - **Xiao et al. (2024)**: *BGE-Reranker: A Packed Resource for General Chinese Embeddings*. [Paper](https://arxiv.org/abs/2309.07597)
        - **Günther et al. (2023)**: *Jina Embeddings: A Novel Set of High-Performance Sentence Embedding Models*. [Paper](https://arxiv.org/abs/2307.11224)
        - **mGTE (2024)**: *Generalized Long-Context Text Representation for Multilingual Retrieval*. [Paper](https://arxiv.org/abs/2407.19669)
        - **Martin et al. (2019)**: *CamemBERT: A Tasty French Language Model*. [Paper](https://arxiv.org/abs/1911.03894)
        - **BCEmbedding (2023)**: *Bilingual and Cross-lingual Embeddings for RAG*. [Paper](https://github.com/netease-youdao/BCEmbedding)

    Attributes:
        method (str): The name of the reranking method.
        model_name (str): The name or path to the **pretrained reranking model**.
        device (torch.device): The computation device (**CPU/GPU**).
        dtype (torch.dtype): The data type for model inference.
        tokenizer (AutoTokenizer): The tokenizer for encoding queries and documents.
        batch_size (int): The batch size for efficient inference.

    Example:
        ```python
        from rankify.dataset.dataset import Document, Question, Context
        from rankify.models.reranking import Reranking

        # Define a query and contexts
        question = Question("What are the benefits of deep learning?")
        contexts = [
            Context(text="Deep learning allows models to extract hierarchical representations.", id=0),
            Context(text="Machine learning includes both supervised and unsupervised learning.", id=1),
            Context(text="Neural networks are a fundamental component of deep learning.", id=2),
        ]
        document = Document(question=question, contexts=contexts)

        # Initialize Transformer Ranker (e.g., BGE-Reranker)
        model = Reranking(method='transformer_ranker', model_name='mxbai-rerank-xsmall')
        model.rank([document])

        # Print reordered contexts
        print("Reordered Contexts:")
        for context in document.reorder_contexts:
            print(context.text)
        ```
    """

    def __init__(self, method = None, model_name = None, api_key = None, **kwargs):
        """
        Initializes **TransformerRanker** for reranking tasks.

        Args:
            method (str, optional): The reranking method name.
            model_name (str): The name or path to the **pretrained reranker model**.
            api_key (str, optional): API key if required (default: None).
            **kwargs: Additional parameters:
                - batch_size (int, optional): Batch size for inference (default: `16`).
                - device (str, optional): Device (`"cpu"`, `"cuda"`, `"auto"`). Default is `"cuda"`.
                - dtype (torch.dtype, optional): Data type for inference (`torch.float32` or `torch.bfloat16`).
        """
        device = kwargs.get("device", "cuda")
        self.device = get_device(device)
        self.dtype = get_dtype( kwargs.get("dtype", torch.float32), self.device)
        self.model = AutoModelForSequenceClassification.from_pretrained(
            model_name,
            num_labels=1, 
            trust_remote_code=True,
            torch_dtype=self.dtype
            ).to(self.device)
        self.model.eval()

        self.tokenizer = AutoTokenizer.from_pretrained(model_name)
        self.batch_size = kwargs.get("batch_size", 16)

    def tokenize(self, inputs: Union[str,List[str], List[Tuple[str, str]]]):
        """
        Tokenizes **queries and documents** for model input.

        Args:
            inputs (Union[str, List[str], List[Tuple[str, str]]]): 
                Query-document pairs to tokenize.

        Returns:
            dict: Tokenized inputs suitable for the reranking model.
        """
        return self.tokenizer(inputs,return_tensors="pt", padding=True, truncation=True).to(self.device)

    @torch.no_grad()
    def rank(self, documents: list[Document])-> List[Document]:
        """
        Reranks a list of **Document** instances using a **Transformer-based Reranker**.

        Args:
            documents (List[Document]): A list of **Document** instances to rerank.

        Returns:
            List[Document]: The reranked list of **Documents** with updated `reorder_contexts`.
        """
        for document in tqdm(documents, desc="Reranking Documents"):
            context_copy= copy.deepcopy(document.contexts)

            inputs = [(document.question.question, context.text) for context in document.contexts]

            batched_inputs = [
                inputs[i:i+self.batch_size] for i in range(0,len(inputs),self.batch_size)
            ]

            scores = []

            for batch in batched_inputs:
                tokenized_inputs = self.tokenize(batch)
                batch_scores = self.model(**tokenized_inputs).logits.squeeze()
                batch_scores = batch_scores.detach().cpu().numpy().tolist()

                if isinstance(batch_scores, float):
                    scores.append(batch_scores)
                else:
                    scores.extend(batch_scores)
            for score, context in zip(scores,context_copy):
                context.score = score
            context_copy.sort(key=lambda x:x.score, reverse=True)
            document.reorder_contexts = context_copy    
        return documents

__init__(method=None, model_name=None, api_key=None, **kwargs)

Initializes TransformerRanker for reranking tasks.

Parameters:

Name	Type	Description	Default
method	str	The reranking method name.	None
model_name	str	The name or path to the pretrained reranker model.	None
api_key	str	API key if required (default: None).	None
**kwargs		Additional parameters: - batch_size (int, optional): Batch size for inference (default: 16). - device (str, optional): Device ("cpu", "cuda", "auto"). Default is "cuda". - dtype (torch.dtype, optional): Data type for inference (torch.float32 or torch.bfloat16).	{}

Source code in rankify/models/transformer_ranker.py

def __init__(self, method = None, model_name = None, api_key = None, **kwargs):
    """
    Initializes **TransformerRanker** for reranking tasks.

    Args:
        method (str, optional): The reranking method name.
        model_name (str): The name or path to the **pretrained reranker model**.
        api_key (str, optional): API key if required (default: None).
        **kwargs: Additional parameters:
            - batch_size (int, optional): Batch size for inference (default: `16`).
            - device (str, optional): Device (`"cpu"`, `"cuda"`, `"auto"`). Default is `"cuda"`.
            - dtype (torch.dtype, optional): Data type for inference (`torch.float32` or `torch.bfloat16`).
    """
    device = kwargs.get("device", "cuda")
    self.device = get_device(device)
    self.dtype = get_dtype( kwargs.get("dtype", torch.float32), self.device)
    self.model = AutoModelForSequenceClassification.from_pretrained(
        model_name,
        num_labels=1, 
        trust_remote_code=True,
        torch_dtype=self.dtype
        ).to(self.device)
    self.model.eval()

    self.tokenizer = AutoTokenizer.from_pretrained(model_name)
    self.batch_size = kwargs.get("batch_size", 16)

tokenize(inputs)

Tokenizes queries and documents for model input.

Parameters:

Name	Type	Description	Default
inputs	Union[str, List[str], List[Tuple[str, str]]]	Query-document pairs to tokenize.	required

Returns:

Name	Type	Description
dict		Tokenized inputs suitable for the reranking model.

Source code in rankify/models/transformer_ranker.py

def tokenize(self, inputs: Union[str,List[str], List[Tuple[str, str]]]):
    """
    Tokenizes **queries and documents** for model input.

    Args:
        inputs (Union[str, List[str], List[Tuple[str, str]]]): 
            Query-document pairs to tokenize.

    Returns:
        dict: Tokenized inputs suitable for the reranking model.
    """
    return self.tokenizer(inputs,return_tensors="pt", padding=True, truncation=True).to(self.device)

rank(documents)

Reranks a list of Document instances using a Transformer-based Reranker.

Parameters:

Name	Type	Description	Default
documents	List[Document]	A list of Document instances to rerank.	required

Returns:

Type	Description
List[Document]	List[Document]: The reranked list of Documents with updated reorder_contexts.

Source code in rankify/models/transformer_ranker.py

@torch.no_grad()
def rank(self, documents: list[Document])-> List[Document]:
    """
    Reranks a list of **Document** instances using a **Transformer-based Reranker**.

    Args:
        documents (List[Document]): A list of **Document** instances to rerank.

    Returns:
        List[Document]: The reranked list of **Documents** with updated `reorder_contexts`.
    """
    for document in tqdm(documents, desc="Reranking Documents"):
        context_copy= copy.deepcopy(document.contexts)

        inputs = [(document.question.question, context.text) for context in document.contexts]

        batched_inputs = [
            inputs[i:i+self.batch_size] for i in range(0,len(inputs),self.batch_size)
        ]

        scores = []

        for batch in batched_inputs:
            tokenized_inputs = self.tokenize(batch)
            batch_scores = self.model(**tokenized_inputs).logits.squeeze()
            batch_scores = batch_scores.detach().cpu().numpy().tolist()

            if isinstance(batch_scores, float):
                scores.append(batch_scores)
            else:
                scores.extend(batch_scores)
        for score, context in zip(scores,context_copy):
            context.score = score
        context_copy.sort(key=lambda x:x.score, reverse=True)
        document.reorder_contexts = context_copy    
    return documents

get_device(device, no_mps=False)

Source code in rankify/utils/helper.py

def get_device(
        device: Optional[Union[str, torch.device]],
        no_mps: bool = False,
    ) -> Union[str, torch.device]:
        if not device:
            if torch.cuda.is_available():
                device = "cuda"
            elif torch.backends.mps.is_available() and not no_mps:
                device = "mps"
            else:
                device = "cpu"
        return device

get_dtype(dtype, device, verbose=1)

Source code in rankify/utils/helper.py

def get_dtype(
        dtype: Optional[Union[str, torch.dtype]],
        device: Optional[Union[str, torch.device]],
        verbose: int = 1,
    ) -> torch.dtype:
        if dtype is None:
            print("No dtype set")
        if device == "cpu":
            dtype = torch.float32
        if not isinstance(dtype, torch.dtype):
            if dtype == "fp16" or "float16":
                dtype = torch.float16
            elif dtype == "bf16" or "bfloat16":
                dtype = torch.bfloat16
            else:
                dtype = torch.float32
        return dtype